Re: [Qemu-devel] [QEMU PATCH v5 1/6] migration: alternative way to set instance_id in SaveStateEntry

[Michael Roth]

Quoting David Gibson (2016-10-10 00:31:20)
> On Fri, Oct 07, 2016 at 09:07:49AM +0100, Dr. David Alan Gilbert wrote:
> > * David Gibson (address@hidden) wrote:
> > > On Wed, Oct 05, 2016 at 09:44:57AM -0700, Jianjun Duan wrote:
> > > > Please see comments below:
> > > > 
> > > > On 10/05/2016 03:12 AM, Dr. David Alan Gilbert wrote:
> > > > > * Jianjun Duan (address@hidden) wrote:
> > > > >> In QOM(QEMU Object Model) migrated objects are identified with 
> > > > >> instance_id
> > > > >> which is calculated automatically using their path in the QOM 
> > > > >> composition
> > > > >> tree. For some objects, this path could change from source to target 
> > > > >> in
> > > > >> migration. To migrate such objects, we need to make sure the 
> > > > >> instance_id does
> > > > >> not change from source to target. We add a hook in DeviceClass to do 
> > > > >> customized
> > > > >> instance_id calculation in such cases.
> > > > > 
> > > > > Can you explain a bit about why the path changes from source to 
> > > > > destination;
> > > > > the path here should be a feature of the guest state not the host, 
> > > > > and so I
> > > > > don't understand why it changes.
> > > > Please see the discussion with David in the previous versions:
> > > > http://lists.nongnu.org/archive/html/qemu-ppc/2016-06/msg00062.html
> > > 
> > > Um.. your description above really isn't an accurate summary of that
> > > discussion.
> > > 
> > > The point is not that the qom path will vary from source to
> > > destination for some arbitrary reason, but rather that we anticipate
> > > future changes in the QOM structure.  Specifically we're considering
> > > eliminating the DRC objects, and folding their (limited) state into an
> > > array in the parent object (either the machine or a PCI host bridge).
> > > 
> > > That would change the qom paths, and hence the auto-generated instance
> > > ids, which would break migration between qemu versions before and
> > > after the restructure.
> > > 
> > > I'm not sure that changing the instance ids is enough though, anyway,
> > > since we're talking about eliminating the object entirely, the
> > > class/type information in the migration stream also wouldn't match.
> > > 
> > > Dave, if you have ideas on how to deal with that, I'd love to hear
> > > them
> > 
> > Eliminating the object entirely would be tricky to deal with;
> > allowing the structure to change would work if instead of a custom 
> > instance_id
> > you had a custom idstr.
> 
> Sorry, two misunderstandings here.
> 
>   * When I said "structure" I meant in the high-level sense of what
>     objects exist and are responsible for what things, not in the
>     'struct WhateverState' sense.
> 
>   * In fact right now eliminating the objects would be ok, since they
>     have no migration state (which causes the problems this series is
>     trying to address).  But applying this series which adds migration
>     state would make it difficult to eliminate the objects in future.
>     That's an awkward constraint given that we've already got some
>     hints that these objects are not a good idea.
> 
> > However, the question then becomes why is the structure changing so much;
> > ideally things in the migration stream should represent some tangible object
> > that corresponds to something real, but (again ideally) the contents
> > of the stream should reflect the state of those objects not the current
> > implementation - so if you want to change the implementation the stream
> > doesn't change.  Is it your implementation, or the understanding of what
> > the objects actually represent that's in flux?
> 
> So, the objects in question are DRCs - "Dynamic Re-configuration
> Connector"; silly IBM talk for "a port into which something can be
> hotplugged", bascially.  These aren't "real" devices, but rather a
> firmware/hypervisor abstraction which are used to describe a hotplug
> point.  Each DRC allows either one CPU core, one PCI device, or one
> LMB (256MiB chunk of RAM) to be hotplugged (or removed).  The PCI DRCs
> are "owned" by the PCI host bridge to which the device would be
> connected, the CPU and memory DRCs are owned by the machine.
> 
> The state variables which Jianjun Duan is adding to migration are
> values defined in the PAPR (hypervisor interface) spec, and so are
> tangible enough to be sensible to migrate.  At the moment, each LMB is
> represented by a discrete QOM object, but I've been thinking for a
> while that this may be a mistake.  In particular it's a problem for
> the LMB DRCs - because each LMB is only 256MiB of memory, we end up
> with thousands, maybe tens of thousands of DRC objects on a guest with
> with large maxmem (even if initial memory is small).  The QOM
> infrastructure doesn't deal terribly well that many object child
> properties.
> 
> The construction of those DRCs used to be an N^3 algorithm, which as
> you'd imagine caused horrible startup times with large maxmem (minutes
> to hours above ~512GiB).  We fixed it to be only N^2, so now the
> startup delays are merely bad (at least once you get to ~2TiB+
> maxmem).
> 
> We could fix that by changing QOM to use a hash or tree for object
> children, instead of a plain list, but I'm not sure if it makes sense
> to do that with only this use case.

Is that still as much an issue as of this patch?

commit b604a854e843505007c59d68112c654556102a20
Author: Pavel Fedin <address@hidden>
Date:   Tue Oct 13 13:37:45 2015 +0100

    qom: Replace object property list with GHashTable
    
    ARM GICv3 systems with large number of CPUs create lots of IRQ pins. Since
    every pin is represented as a property, number of these properties becomes
    very large. Every property add first makes sure there's no duplicates.
    Traversing the list becomes very slow, therefore QEMU initialization takes
    significant time (several seconds for e. g. 16 CPUs).
    
    This patch replaces list with GHashTable, making lookup very fast. The only
    drawback is that object_child_foreach() and object_child_foreach_recursive()
    cannot add or remove properties during traversal, since GHashTableIter does
    not have modify-safe version. However, the code seems not to modify objects
    via these functions.

> 
> So, for some time, I've been considering rewriting the DRC stuff to be
> implemented via QOM interface on the "owner" object, rather than
> separate objects for every single connector.

At a high-level I think I'd prefer improving QOM over working around it's
performance issues. It seems like those limits are bound to be pushed
eventually given the wide range of hardware it's meant to support
in theory.

But identifying a DRC state in the migration stream with a stable value based
on DRC index makes sense either way. If I understand this patch it seems
like we have that: if a custom instance_id getter is available, we avoid
using the QOM path for idstr and use the default of vmsd->name, then use
the instance_id from the getter. If we dropped DRCs as objects, we'd need
a mechanism other than DeviceClass->dev_get_instance_id() to provide this
custom instance_id, but otherwise so long as the state was the same, and
the VMSD was the same, it seems like it should be doable in theory.

> 
> So the difficulty is that if we add the state migration to the DRC
> objects as they stand, we prevent ourselves from doing a refactor
> that's probably a good idea.  But no-one's yet had time to actually
> thrash out such a refactor. :/
> 
> -- 
> David Gibson                    | I'll have my music baroque, and my code
> david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
>                                 | _way_ _around_!
> http://www.ozlabs.org/~dgibson